A physics‐consistent transfer learning framework for small‐sample green ammonia yield prediction under operational drift

Abstract

Green ammonia is an important zero‐carbon energy carrier to achieve deep decarbonization. Its production process is highly coupled with intermittent renewable energy, which leads to fluctuations in feed condition fluctuations and a scarcity of steady‐state operational data, bringing severe challenges to traditional data‐driven modelling. In this paper, we transform the green ammonia modelling problem into a domain shift task in transfer learning, and propose a mechanism‐guided interpretable transfer learning framework. Firstly, based on the process simulation software (UniSim Design), a Haber‐Bosch mechanism model covering a wide range of operating conditions is constructed to generate high‐fidelity source domain data, a synthetic dataset of 100,000 operating points (covering 380–450°C, 10–20 bar, and varied H ₂ /N ₂ ratios), and a deep neural network is pre‐trained to encode general nonlinear dynamics patterns. Subsequently, on the target domain with only a small amount of real operation data, the backbone network is frozen to preserve physical consistency, and only the lightweight linear output head is fine‐tuned to calibrate the operating condition bias. This design not only significantly reduces the dependence on field annotation data, but also verifies that the response logic of the model on key operational variables is highly consistent with chemical principles through interpretability analysis. Experiments show that the proposed method only needs 100 target domain training samples to achieve a good prediction effect ( R ²  = 0.9511, MSE = 2.4143 (t/h) ² ), and reaches performance saturation at 25% labelling ratio, which is better than pure data‐driven model and end‐to‐end fine‐tuning strategy. This work provides an intelligent modelling paradigm with high accuracy, strong robustness, and engineering confidence for data‐scarce zero‐carbon chemical processes.